Raw observations
Video, FMG, IMU, tactile and robot logs
DATA PLATFORM & OPERATIONS
From capture logs to reusable data assets.
Multi-device ingestion, time alignment, cleaning and annotation, QA and dataset versioning in one workflow, for lab research and enterprise-scale production alike.
FIELD / REAL-WORLD DATA COLLECTION
FieldCapture App
Turn every real-world manipulation into usable data.
A mobile workflow from brief to recording for research teams, enterprise capture projects and data contributors: create and accept tasks, capture and submit, and get quality feedback.
Turn capture goals into clear tasks.
- 01Define tasks
Describe the manipulation goal, scene, devices and what counts as a valid episode.
- 02Publish and assign
Set who can take part, capture instructions and submission rules.
- 03Track submissions
Follow capture progress and submission records per task.
- 04Review and re-capture
Turn quality feedback into correction or re-capture tasks.
FROM TASK TO DATA
GET FIELD
Start with one task.
Download coming soon. Supported systems, devices and features are subject to the released version.
DATA CENTER / THE PROJECT VIEW
Data Dashboard.
From a single record to the whole project.
Open the standalone data platform to see capture volume and task and scene coverage, and browse curated episodes. Field teams see their progress; data managers get a solid foundation.
- Data overviewAggregates capture volume and quality metrics to show how your data is growing.
- Task mapOrganizes records by task and scene to show data coverage.
- Data marketplaceBrowse and preview published episodes by scene, task or tag.
Standalone platform · Opens in a new window · Sign in with an authorized account; data visibility follows account permissions.
DATA / CONNECTEDData structure
Tasks and scenesOne shared context
Visual recordsInertial recordsTask labels
Quality reviewEpisode browsing
View data by project.
Sign in to the data platform for actual records and statistics.
- 01
Ingest and sync
Device config, raw recordings, clocks and calibration
- 02
Processing and QA
Cleaning, segmentation, annotation, anomaly review
- 03
Dataset release
Versions, rights, splits, export manifests
- 04
Training & Evaluation
Environment, weights, metrics and real-robot logs
- 05
Failures and re-capture
Locate failure scenarios and add corrective demonstrations
Close the loop: get the smallest task working first, then let evaluation results decide what to capture next. Each step is configured and validated per project; nothing is automated by default.
Layered data, trustworthy results.
Each episode links task, device, time and calibration. Raw signals stay traceable; algorithm outputs keep their source and version and never overwrite the original record.
Calibration and estimation
Time alignment, coordinate transforms, hand or tool poses
Annotations and labels
Task segments, events, human review, robot actions
Dataset versions
Train/validation/test splits, export manifests and licenses
SIGNAL VOCABULARY · DATA ENGINE
How we measure the physical world.
Each modality is a way of capturing or understanding real operations. Together they form the signal vocabulary behind the dataset — from egocentric vision and contact signals to sync, calibration, and quality gates.
Pixel VisionPV-01 · PH-SIG
Depth & GeometryDG-02 · PH-SIG
Segmentation & KeypointsSK-03 · PH-SIG
FMG ContactFM-04 · PH-SIG
Tactile ForceTF-05 · PH-SIG
IMU & MotionIM-06 · PH-SIG
Point Cloud & 3DPC-07 · PH-SIG
Time SyncSC-08 · PH-SIG
CalibrationCB-09 · PH-SIG
Quality RulesQR-10 · PH-SIG
These are signal dimensions commonly used in Physen capture and platform work. Fields, sampling, and error budgets follow the project protocol and device revision; no modality is assumed to improve task success by default.
A SHARED DATA CONTRACT
Agree on conventions first, then connect devices and algorithms.
Recommended data standards for projects. Actual fields, format adaptation and platform features depend on the selected version and pilot validation.
Identity and versions
Project, task, episode and device IDs, plus firmware, calibration, algorithm and data versions.
Time and space
Device time and reference clock, sync method and error; coordinate frames, units, extrinsics and validity windows.
Quality and provenance
Dropped frames, gaps, anomalies and human review results; measured values, estimates and human labels kept distinct.
Rights and splits
Capture consent, scope of use and access rights; evaluation data isolated by person, scene or date.
SCALE COLLECTION, NOT UNCERTAINTY
Data delivery at scale, with clear units of acceptance.
We start with a small pilot to calibrate quality and processing cost, then plan people, devices, scenes and batches. Delivery can cover some or all of capture, processing, QA and dataset building.
What each batch delivers
- Raw and processed data, with field dictionaries and reading guides
- File manifests, version IDs, checksums and calibration files
- Valid-episode statistics, coverage distribution, QA and spot-check reports
- Rejection reasons, rework records and usage-rights notes
What we confirm before kickoff
- How valid episodes, duration or completed-task records are counted
- Quality thresholds such as time offset, missing rate and annotation consistency
- Coverage requirements for tasks, objects, environments, participants and failure cases
- Ownership, storage region, retention period and third-party processing rights
Let failures define the next batch
Use evaluation logs to locate failure states and build a re-capture list. Corrective demonstrations go into a separate version and are compared against a fixed test set. Broad coverage is kept, so capture does not just chase the current model's local failures.
Connect your own devices and third-party algorithms
Interfaces, output conventions, permissions and adaptation boundaries.
Connect to training, not just storage
Link data versions to training configs and evaluation results.
START WITH YOUR TASK
Tell us about your research task.
From device selection to data and training workflows, we'll help you find the right place to start.